Alt+tab is an optimal controller.
Maybe operating on the order of items in the queue, and use your prediction to sort windows, allowing faster selection? Even that disrupts sense of place - I know what applications I have open and where they are, and if I'm using alt+tab over 5 or more, I know the order in which I've opened them, and "where" I need to go to navigate to them.
There are second and third order impacts to changing interface behaviors, so the superficial benefit you might gain will be lost by creating friction at different levels.
A single key is insufficient for granular control, and no AI widget short of human level AI is going to capture the edge cases, which will create friction, at which point I will aggressively remove the offending piece of software.
I'd go back to the drawing board and work on a more complex model of window switching and all the ways in which people use alt+tab, and see if there's a use case for your idea at a different level. As it is, for me, it would interfere with a reliable and predictable interface, and I would be very unhappy.